Highlights from the 28th Annual Meeting of the Society for the Neural Control of Movement

INTRODUCTION

From April 30th to May 4th, 2018 the Society for the Neural Control of Movement gathered for their 28th annual meeting in Santa Fe, New Mexico. Over 280 scientists, students, and clinician-investigators discussed recent developments in theoretical as well as experimental motor control research and potential clinical applications. Throughout the meeting, diverse panels discussed topics ranging from “how musculoskeletal mechanics relate to proprioception” to “how health and disease affect motor learning.” Preceding the annual meeting, a satellite meeting was hosted by the Santa Fe Institute entitled “The Complexity of the Nervous System.” The society increased its effort to foster diversity and open intellectual exchange with a focus on adherence to an official code of conduct for a respectful and fair dialog. For the first time in the history of this meeting, one panel consisted entirely of female speakers to discuss how motor control experiments were being tested in real-world environments. Additionally, the society provided financial support for researchers from developing countries and promising young scientists.

This editorial describes several research highlights emphasized at the meeting. A word cloud depiction of the most talked about topics presented at the meeting is shown in Fig. 1. For the purposes of this review, the research presented was categorized into four topics: 1) approaches to move motor goal research toward real-world conditions, 2) considerations to develop effective mathematical models, 3) recent developments in motor learning and brain plasticity, and 4) work on integrating sensory information for motor control. Highlights from each of these topics describe the current direction of experimental research for the neural control of movement.

Fig. 1.

The most common keywords reported by attendees of the Neural Control of Movement meeting representing what was heard or learned.

MOVING MOTOR-CONTROL RELATED RESEARCH INTO REAL-WORLD CONDITIONS

Transitioning from “simple” movement experiments to more complex tasks that would be experienced in everyday life was an important point of discussion throughout the meeting. How do we extend our understanding of the neural principles underlying motor control gained from simple movement experiments (such as reaching and grasping) to explain more complex and arguably more naturalistic movements (such as walking with a cup full of coffee or delicately peeling a piece of fruit)? To begin building these bridges, members of the society have broadly taken three synergistic approaches: 1) constraining “real-world” behaviors to make them quantifiable and controlled, 2) loosening the constraints on well-defined laboratory-based paradigms either by allowing more degrees of freedom in movement or modalities or by taking into account physiological parameters, and 3) creating new tools and sophisticated behavioral analyses that can capture the complexity of the rich behavioral output.

A variety of innovative behavioral paradigms were presented to approach the diversity of movements humans ordinarily perform. Virginia Penhune studied the acquisition and performance of the skills required to play music in a quantifiable and experimentally controlled environment. Penhune utilized novel fiber optic-based magnetic resonance scanner-compatible musical instruments (5-note cello, 14-key piano) in conjunction with functional magnetic resonance imaging in human subjects to investigate the neural basis for complex multimodal learning. She presented work from her laboratory showing neurophysiological support for the presence of a sensitive period for musical training (Bailey et al. 2014; Steele et al. 2013) based on recent experiments designed to identify the roles of different modalities (visual, auditory, or motor) in musical training. Trained musicians learned from auditory cues better than from motor cues, whereas untrained musicians learned equally well from either modality, suggesting musical training improves translating auditory commands into action.

A complementary approach for studying more naturalistic behavior is to make the existing laboratory paradigms richer by loosening experimental constraints while carefully introducing tractable, complex elements. Dagmar Sternad presented a series of studies focused on determining the strategies associated with the control of complex objects. She described a behavioral paradigm with a haptic interface that implements a virtual reality-based cart-and-pendulum model, loosely mimicking a task that requires moving a cup full of coffee without spilling (Hasson et al. 2012). This behavioral paradigm extended the question of neural control of movement from discrete point-to-point reaches in two important ways. First, using an object that is not rigid introduced a chaotic component during the physical interaction for which the motor system has to account. Second, continuous interaction with the object gives a better window into understanding the underlying temporal dynamics of the control strategy. Using this behavioral paradigm, Sternad and colleagues investigated the strategies subjects adopted to not “spill the coffee” (i.e., to not let the pendulum leave the cart under varying conditions). Perturbations with speed bumps and constraints on movement amplitude were presented to the subjects to characterize the degree to which the coffee cup was controlled. Posters from Sternad’s group presented data from a novel paradigm for which subjects move a cup holding a ball horizontally on a 2D surface while their motion is captured by an overhead camera. Critically, they found humans did not prioritize interaction force (a proxy for effort), but rather strategies that rendered interactions predictable (Maurice et al. 2018).

In a similar effort to transition motor experiments to more real-world scenarios, Gelsy Torres-Oviedo presented wheeled motorized shoes whose gain could be independently controlled as an alternative to split belt training. She showed that subjects were able to adapt to an asymmetric gain on the shoes with adaptation persisting when subjects were tested on the ground, thus showing evidence for generalization. While regular practice on a treadmill leads to long-lasting correction of step length asymmetry poststroke, this correction does not always generalize well to naturalistic overground conditions. With the use of motorized shoes, Torres-Oviedo was able to experimentally test how changes in the natural environment affected walking characteristics. The difference between the laboratory setting and a naturalistic training environment is the lack of visual flow and control over speed and step initiation.

How does the brain estimate the cost of movement and make a decision on which strategy to choose? Optimal feedback theory, an influential theory for the control of movements, relies on cost optimization minimizing movement effort (Todorov and Jordan 2002). While the contribution of metabolic costs to movement effort is under debate (Morel et al. 2017), Alaa Ahmed presented a series of experiments studying effortful reaching movements in healthy humans while using respirometry to measure metabolic costs. When subjects performed reaching movements against with a fix load, metabolic cost rather than squared rate of torque (Uno et al. 1989) or squared muscle force (Schweighofer et al. 2015) was the best predictor for effort (Shadmehr et al. 2016). Given effort has an effect on movement speed, Ahmed showed the metabolic cost model can be effectively used to explain the self-selected speed of participants when performing reaching movements with different loads. Eric Summerside presented his findings that metabolic cost is not affected by age; however, movement speed is affected by reward and age, which suggests a change in subjective rather than objective value of effort.

Surgery is one of the most dexterous skills that humans can perform. While quantification of complex natural skills remains fairly elusive, teleoperated robot-assisted minimally invasive surgery provides an exciting opportunity to study the evolution of dexterity. Ilana Nisky and her colleagues utilized a teleoperated surgical system (the da Vinci Research Kit/da Vinci Surgical System) to analyze the movements of expert surgeons and novice users (Sharon and Nisky 2017). First, they found movements performed by the operators with the surgical robot follow movement laws observed in drawing movements, such as the two-third power law (or the one-sixth power law in 3 dimensions). Second, when they characterized the parameters of the power law, they found a clear difference between expert surgeons and nonexpert users in parameters that were specifically related to the tempo and the smoothness of movements. Furthermore, these parameters were strongly affected by the movement segment (i.e., transport of needle, insertion, position of the needle, etc.). Importantly, the movement kinematics of surgical planning captured by this system were amenable to movement segmentation algorithms, which could lead to insight into the process underlying the construction of skilled movement sequences.

Similarly, behavioral paradigms in animal models can be expanded to allow for the expression of a larger behavioral repertoire. Michael Berger presented a novel approach for investigating cortical activity in unrestrained nonhuman primates. By developing an experimental environment within a monkey cage, it is now possible to study goal-directed behavior involving walking and reaching of rhesus monkeys while recording intracortical single unit activity (Berger and Gail 2018). Berger investigated activity in the fronto-parietal network regarding movement planning for reaching movements to near-located and walk-and-reaching movements to far-located movement goals. While evidence from neglect patients suggests different brain areas could be responsible for processing the space near and far from the body (e.g., Halligan and Marshall 1991), Berger presented results showed planning activity is partly modulated independent of distance.

To leverage the strength of a model system, existing behavioral paradigms can be adapted to other species. Tejapratap Bollu presented a novel real-time home-cage joystick system that can train mice on a center-out reaching task, typically studied in primates (Bollu et al. 2018). This system automatically trained mice to produce complex, directed center-out forelimb trajectories while implementing closed-loop optogenetics. By using a high-throughput system (i.e., large sample sizes), Bollu and colleagues were able to rapidly test multiple hypotheses of how different motor cortical areas contribute to maintaining limb position and reaching to targets. With trajectory decomposition methods they found inactivation of the contralateral caudal forelimb area in mice reduced the peak speed of kinematic primitives but preserved their timing, complexity, and direction. As a result, the mouse trajectories exhibited isomorphic hypometria, i.e., they remained appropriately directed to targets but were spatially contracted.

Several posters at the meeting presented methods helping to investigate more natural conditions. For unstructured nonrepetitive behavior, it is difficult to identify the underlying dynamics in the neural data as usually repetition is used to separate the latent signal from noise. A deep learning approach is the Latent Factor Analysis via Dynamical Systems (LFADS) present by the posters of Chethan Pandarinath and Lahiru Wimalasena. Once trained, LFADS essentially “denoises” single-trial neural data and captures only the relevant dynamics. This denoised signal can be used to predict behavioral data, for instance reach kinematics, with superior performance compared with other methods such as GPFA or Gaussian smoothing (Pandarinath et al. 2017). Wimalasena applied this method to motor cortical activity of a monkey performing an unstructured reaching tasked for which the monkey was asked to reach to a random position on screen. Linear reconstruction of arm muscle potentials for single trials was significantly better when using the LFADS signal compared with the Gaussian smoothed spike trains.

Mackenzie Mathis presented a promising approach based on a deep learning architecture for markerless tracking across species (Mathis et al. 2018). Due to its utilization of transfer learning in deep artificial neural networks (Yosinski et al. 2014), the network was trained to human-expert level performance with a few hundred manually coded frames even under challenging lighting conditions.

Andrea Pack presented two novel electrode systems, one using carbon nanotube fibers and another a multichannel flexible electrode system that could record single-unit EMGs from the syrinx of singing birds, bringing us closer to understanding the interaction between nervous system and muscle underlying the generation of precise movements. These studies show that in addition to studying complex movements, broadening experimental paradigms to ask similar questions across species is also likely to identify and elucidate the role of the underlying neural circuits in behavior that might not immediately be evident in humans or nonhuman primates.

Overall, these presentations highlighted the need for new methodologies to study less constrained behavior. Investigating neurophysiological correlates of free behavior requires addressing at least three major difficulties: 1) reliable measurements of behavioral parameters in unconstrained experimental environments, 2) sophisticated behavioral analysis that captures the complexity of the behavioral data, and 3) neurophysiological analysis of unstructured nonrepetitive neural data. Generalizing experimental findings from simplistic tasks in the laboratory to complex real-world situations remains challenging; however, the focus of the meeting on this topic showed real progress was being made to bridge this gap.

DEVELOPING EFFECTIVE MATHEMATICAL MODELS OF MOTOR CONTROL

When attempting to understand motor control in the real world, a core problem for any scientific design is to develop useful models that provide effective descriptions. This was addressed during the satellite meeting before the conference, organized by David Krakauer and colleagues from the Santa Fe Institute. The institute studies complex adaptive systems to develop universal theories that can be applied to a variety of scientific disciplines, including neuroscience. In his opening talk, David Krakauer introduced the idea of a spectrum between completely ordered dynamics (like the physics of billiards) and completely random dynamics (like the physics of heat transfer). Complexity lies in the middle of this spectrum, where the dynamics of a system cannot be boiled down to simple explanation (e.g., evolution by natural selection). These complex systems present an interesting problem: unlike with systems at the extreme ends of this spectrum, the ideal level of abstraction at which to study complex systems is often unclear.

This problem is particularly salient for studying how the brain generates behavior. As John Krakauer’s recent paper (Krakauer et al. 2017) points out, there exist incredible tools for studying the brain. Unfortunately, the proliferation of these tools can often cause a reductionist bias in studying the brain. As with all complex systems, an effective theory of the brain must be based on studying the brain at relevant levels of abstraction to the question being asked. As an extreme example, if the experimental question involves asking how does the brain generate reaching behavior, it makes little sense to start by studying the quantum mechanics of ion channels. Instead, a useful abstraction for this level of question might be to model certain regions of the brain as a network of neurons, communicating specifically to generate a reach.

In the vein of developing effective theories for complex networks, Michelle Girvan showed there is a tradeoff between the responsiveness of a network to its inputs and the stability of its dynamics. In large networks, due to a large set of interacting nonlinearities, the dynamics of a network’s response can easily devolve into chaos, for which small changes in the input to the network result in wildly different outputs. When a network is trained, by restructuring its connections, to generate a robust input-output relationship, these chaotic dynamics are curbed in favor of more stable dynamics. However, Girvan showed that if these dynamics are curbed too strongly the network has a difficult time generating a rich input-output relationship. Thus, it seems networks are often structured such that they are at a critical point, on the brink of chaotic dynamics, to have maximal expressibility (Lu et al. 2016). As another example of finding coarse-grained effective variables to explain behavior of complex system, Geoffrey West presented a theoretical development of simple power law between metabolic energy and body mass which revealed that necessity of sleep is mainly for reorganization and repair of the brain, and not the body (Savage and West 2007). Both Girvan's and West’s approaches demonstrated how an effective theory can extract low-dimensional and simple principles that govern a complex behavior which, at surface, involves high-dimensional space of parameters and requires intense computations to describe.

Artemy Kolchinsky discussed neural networks, which not only emerged as a subfield in complexity science but recently “regained” popularity in neuroscience as a computational tool to capture latent, abstract features underlying behavior of activities in network of neurons. When neural networks were first introduced in the 1980s, they generally did not outperform other supervised learning algorithms. Since then, deep neural networks were shown to outperform many other algorithms and even human experts in various applications like image recognition (Krizhevsky et al. 2012), language translation (Sutskever et al. 2014), and strategic board games like Go (Silver et al. 2017). While the structure of artificial neural networks is loosely inspired by the biological behavior of networks of neurons, deep neural networks are designed for high performance in their specific task rather than biological relevance. Deep neural networks differ from conventional neural networks in three aspects that led to the better performance: 1) using more layers (“deep”) of nodes leads to significant improvement in processing information; 2) noise, which is ubiquitous in neural activities, can be an effective regularizer; and 3) structural design of network layers provides a way to formalize inductive bias. Despite performance improvements of the learning algorithms, there remains a gap between the capabilities of deep neural networks and the capabilities of the brain, such as improving the ability to learn from small amounts data, improving pattern recognition in data without labels (i.e., unsupervised learning), enhancing robustness to adversarial inputs, and expanding the currently nascent applications of deep neural networks to learning of natural language, causal reasoning, and motor control.

Finally, Jessica Flack and colleagues’ work toward developing principles of collective computation for adaptive systems provided highly relevant insights to theoretical perspectives in motor control. Flack argued while addition of information from microscopic inputs introduces subjectivity, an adaptive system produces order through collectively estimating regularities to find an output with functional macro property (Flack 2017). To harness diversity in individual entities, Flack and her colleagues (Daniels et al. 2017) further proposed collective computation occurs in two phases: first, an information accumulation phase from which a pattern is detected; and second, an optimization process to best aggregate accumulated information in presence of biases and errors. Flack showed such principles of collective computation can explain social or neural decision-making process.

Interestingly, Flack’s work gave rise to discussions of two actively debated topics in the motor control community. First, as this collective computation involves extracting a coarsely grained representation of microscopic inputs, many in the motor control audience drew a parallel with dimensionality reduction techniques (e.g., principal component analysis or nonnegative matrix factorization) to decompose kinematics (Hogan and Sternad 2012), muscle activity (Giszter 2015; Ting et al. 2015), or neural activity (Cunningham and Yu 2014; Gallego et al. 2017; Sadtler et al. 2014). Classically, these methods have been used to describe movement using low-dimensional, higher-level variables. While neural origin of such low-dimensional control of movement is still controversial (Bizzi and Cheung 2013; Kutch and Valero-Cuevas 2012), Flack’s work provides a theoretical framework that can be useful in testing such hypothesis in a mechanistically principled way.

Second, Flack pointed out a solution found as a consequence of collective computation is not necessarily optimal for any particular goal but can be “good enough” across a broad range of goals. Such findings closely parallel a theoretical perspective in motor control arguing a motor solution a brain or an individual chooses to use is not optimal but good enough, which facilitates generalization across broader contexts (Loeb 2012). While some specific examples of effective theory introduced at this Satellite Meeting were not immediately applicable to brain or motor behavior, it can be thus concluded the rationale behind developing effective theories have substantial implications to studying neural control of movement.

DEVELOPMENTS RELATED TO MOTOR LEARNING AND BRAIN PLASTICITY

To create accurate models of how the brain produces behavior, aspects of how the brain and body undergo plastic changes must be incorporated. Brain and body undergo plastic changes when encountered with challenges that require adaptation of behavior, acquisition of novel skills, and development of compensatory mechanisms to mitigate impairments resulting from neurological injuries. Understanding this process has long been the focus of many members of the society. At this year’s meeting, many presentations shed light on novel insights on sensorimotor learning and plasticity at various systemic levels (e.g., synaptic activity in brain to behavior), at different timescales (e.g., hours to months), and across diverse tasks (e.g., sensory, sequence leaning, simple to complex motor skills).

The simplicity with which motor experiments are designed creates a challenge for identifying how the brain learns to perform a new task or skill. Traditionally, motor control studies require subjects to move a cursor to acquire targets in a two-dimensional workspace (such as a “center-out task”). Adrian Haith presented a novel task in which human participants performed a task that involved bimanual control of a cursor to acquire different targets. This task required the participant to move the left hand up and down to move the cursor left and right, and to move the right hand left or right to move the cursor up and down. Thus, the task involved movements of both hands to control two-dimensional movement of the cursor in a complex manner. Participants showed drastic reductions in movement times and target acquisition accuracy compared with using only one hand to control the cursor. By perturbing the trajectory of the cursor during bimanual control, participants made small corrections in velocity to change the direction; however, no such correction was observed during unimanual control. The complexity of using both hands to navigate the 2D workspace allowed for perturbations to effect motor control in ways for which the simpler unimanual control would compensate.

As individuals learn to perform a task, the nervous system integrates motor control with other afferent signals. Startle-evoked movements (called startReact) provide a measure of how movements can be involuntarily produced in response to a startling stimulus (e.g., loud sound). A wide range of planned movements can be evoked by startReact, ranging from reaching and grasping to postural balance. Claire Honeycutt presented her findings that movements which did not produce a startle response could be trained to do so. Honeycutt described how previous studies of individuated finger movements had not been previously shown to respond to a startle response. This may be due isolated finger movements being rarely performed in daily living even though the movements are fairly easy to perform (Kirkpatrick et al. 2018). Participants performed individuated finger movements over a 2-wk training window and subsequently became susceptible to startReact, which caused those fingers to extend. She recruited “experts” who performed 2 wk of extending and flexing each of their fingers separately with “nonexperts” who did no such training. StartReact was present in the expert group but not for the nonexpert group, suggesting learned motor skills integrate neural circuits in other sensory areas such as spinal reflexes.

Kevin Mazurek’s presentation highlighted how subjects can learn to perform a movement task using intracortical microstimulation instructions delivered to nonsensory areas. Nonhuman primates learned to associate microstimulation delivered to different electrodes in the premotor cortex with performance of different actions in a reach, grasp, and manipulate task (Mazurek and Schieber 2017). Over the course of several weeks, monkeys learned to use the microstimulation as instructions for manipulating different objects with comparable performance to when the instructions were delivered with visual cues (in the form of blue lights). His findings demonstrated monkeys could learn to perform a task when the instructions were delivered using microstimulation in either nonsensory areas or sensory areas. This has implications for neuroprostheses for individuals with focal lesions where cortical regions no longer communicate. By recording neural activity above and using microstimulation to deliver the information below the focal lesion, future neural prostheses could effectively bypass the damaged connections and restore communication. Mazurek’s findings are an important first step toward identifying where in the brain information can be delivered using microstimulation that subjects can learn to use as instructions for performing specific movements.

Frédéric Crevecoeur highlighted the role of rapid somatosensory feedback in guiding the initial phase of learning. Crevecoeur showed that, during early exposure to novel dynamics (e.g., curl force field), subjects modulated their control by increasing movement velocity and upregulated their feedback gains, probed by randomly interleaved step perturbations. Such trial-to-trial modulation can be described by a control strategy known as robust control, which renders the neural controller insensitive to unknown disturbances while trading off efficiency. To further examine how subjects updated their control policy online, Crevecoeur asked his subjects to reach toward visual targets with or without via-point, with or without force field. Adaptation of subjects’ movement velocity after the via-point depended on the type of force field experienced before getting to the via point. Crevecoeur claimed that this reflects an adaptive control, which is an effective strategy to counter potential errors in state estimates (i.e., internal models).

David Franklin demonstrated how the rapid visuomotor pathway is tuned and fractionated to the complexity of environmental dynamics. Franklin showed visual shifts of the hand representation (cursor) during reach produce rapid corrective motor response that is involuntary in nature and is not affected by background loads — i.e., does not exhibit gain scaling as proprioceptive pathway (Franklin et al. 2012, 2016). These rapid visuomotor responses could be upregulated and downregulated according to the visual statistics of the environment and fractionally tuned independent response to perturbation direction according to the environment (Franklin et al. 2014; Franklin and Wolpert 2008). Finally, Franklin demonstrated there are two components of feedback gain changes: reactive and predictive. By comparing adaptation in rapid visuomotor responses to either abrupt or gradual introduction of force field, Franklin showed that reactive feedback gains are excited with increased uncertainty of the predictive feedforward model whereas predictive feedback gains are gradually tuned to the environmental dynamics and statistics (Franklin et al. 2017).

Chao Gu presented a study in which rapid visuomotor responses can be used as a neuromuscular signature of the learning process. Gu measured a rapid stimulus-locked response (SLR) elicited by presentation of a visual target, which is a small burst of muscle activity whose directional tuning is normally aligned to later movement-related activity (Gu et al. 2016). Gu asked whether motor learning alters the directional tuning of the SLR. When subjects adapted their reaching to abruptly introduced visuomotor rotations, there was partial adaptation of SLR, where change in preferred direction relative to prerotation was smaller than what would be considered as full adaptation. To further dissociate whether the change in SLR tuning was to the implicit vs. explicit components of motor learning (Taylor et al. 2014), Gu introduced a gradual visuomotor rotation task from which he found that explicit awareness is not required for SLR adaptation whereas implicit learning influences the SLR tuning. Finally, using a mental visuomotor rotation task (Mazzoni and Krakauer 2006), Gu showed change in explicit aiming does not change SLR tuning. Overall, Gu’s work demonstrated that SLR tuning is selectively modulated by implicit learning. Given extensive projections between the cerebellum and the reticular formation, Gu speculated that such finding provides further evidence that a tecto-reticulospinal pathway mediate the SLR (Corneil and Munoz 2014), because individuals with cerebellar deficit (e.g., cerebellar ataxia) lack ability to utilize implicit learning processes (Taylor et al. 2010).

Aiko Thompson demonstrated a promising utility of training rapid feedback pathways in restoring motor function. Thompson showed that inducing guided changes in spinal reflex pathways through rewarding a stimulus-triggered EMG response (i.e., operant conditioning) improves impaired locomotion after spinal cord injury. More specifically, she introduced an operant conditioning protocol in which subjects were trained to modify their H-reflex amplitude (provided as visual feedback) of soleus muscle during standing over the course of 8 wk. Subjects with spastic hyperreflexia due to incomplete spinal cord injury were able to decrease their H-reflex size after training (e.g., down-conditioning). Successful conditioning was shown to restore a more normal step rhythm, as measured by improvement in walking speed and step symmetry (Thompson et al. 2013). Thompson argued that functional impact of reflex conditioning is far greater than what can be explained by changes in the targeted reflex pathway, but rather reflects an overall beneficial plasticity induced across the complex hierarchy of brain and spinal cord (Thompson and Wolpaw 2014). In effort to investigate the underlying mechanism, Thompson presented evidence from animal study suggesting that cortical sensorimotor rhythm (SMR), as examined by the impact of ongoing sensorimotor cortex activity on the H-reflex, modulated the spinal reflex excitability (Boulay et al. 2015). Based on further observation that size of somatosensory evoked potential (SEP) also change with H-reflex during conditioning, Thompson speculated that short-term task-dependent change in the SEP amplitude, along with the SMR activity, may partly explain change of “reflex sensation” often reported by subjects with central nervous system injury.

Understanding how learning affects the nervous system at the synaptic or neurotransmitter level was another topic of discussion at the meeting. Takeo Watanabe presented his findings on how visual perceptual learning (VPL) performance increased as a result of training and how overlearning might hyperstabilize a skill as reflected by levels of GABA and glutamate (Shibata et al. 2017). VPL is defined as improvement of a visual task as a result of visual learning (Sasaki et al. 2010), and Watanabe presented his findings in which magnetic resonance spectroscopy (MRS) was used to measure concentrations of glutamate and GABA over time as participants performed the task. He presented a series of experiments in which participants performed different training sessions varied in duration to represent the ability to overtrain in the VPL task. Participants were divided into two groups: a “no overlearning” group and an “overlearning” group. Both groups performed four VPL sessions in which a different visual orientation was instructed to be detected in each session, with the overlearning group training for twice the duration of the no overearning group. For the no overlearning group, initial performance decreased after the first session but significantly increased after the second session. For the overlearning group, initial performance increased after both sessions. Watanabe used MRS to study how glutamate and GABA related to the hyperstabilization that occurred during overlearning. By measuring the ratio of glutamate (excitatory, E) to GABA (inhibitory, I) detected using MRS, an E/I ratio was compared for the different learning conditions. For the no overlearning group, the ratio increased initially (within minutes) and then decreased to constant value. For the overlearning group, the ratio rapidly decreased lower than baseline and the returned to stabilize. Overall, his findings suggest overlearning strongly stabilized the learning state and was reflected by the E/I ratio detected using MRS.

David Ostry used MRS to determine how GABA and glutamate concentrations changed over the course of learning. He used an audiomotor task in which subjects were presented a sound and had to move a cursor to a position along a half circle that matched the sound (an audio center-out task). Participants learned which positions along the circle would produce sounds that matched the target tone. Using MRS, GABA was shown to increase between the training sessions by ~10% from the initial scan; however, glutamate did not present any change until the posttraining scan (possibly reflecting a postlearning decline). When comparing the movement errors in the task with GABA concentrations, the degree of error was low when the concentration of GABA was found to be high and vice versa. For glutamate, concentrations were in phase with movement errors of the trial. These results suggest both GABA and glutamate track the time course of movement error, and changes in GABA after learning appear to indicate final retention of the task.

SENSATION WITH RESPECT TO MOTOR CONTROL

The last major topic that appeared in multiple presentations at the meeting was the importance of sensation to the neural control of movement. Many presentations focused on proprioception, the sense of body state. While this sense typically remains below the conscious level, proprioception is critical for the coordinated control of movement. In the panel “Reinterpreting proprioception in terms of musculoskeletal mechanics,” Lena Ting and colleagues presented new takes on classic proprioception literature. Raeed Chowdhury studied proprioceptive neurons in the central nervous system, both in the spinal cord and in the primary somatosensory cortex. Classic literature in both of these areas describes these neurons as being tuned to the state of the limb end point, a rather high-level description of limb state, compared with the signals originating in muscle sensors, like spindles. However, Chowdhury showed that this apparent limb end point representation in both areas arises directly from musculoskeletal geometry of the limb (Chowdhury et al. 2017).

Lena Ting and Kyle Blum followed up by focusing on the function of muscle spindles, one of the main sensors for proprioceptive information. While these sensors are classically considered to transduce muscle length and stretch velocity, Ting and Blum showed evidences that spindles appear to have a strong relationship with force on the intrafusal fibers (Blum et al. 2017). Blum went on to describe a mechanistic spindle model that reproduced many physiological phenomena uncaptured by previous mechanistic models, including a dependence of the spindle firing rate on the history of spindle stretching. Friedl De Groot closed out the panel by describing how a force-based model of muscle spindle activity explained clinical measures of spasticity in spinal cord injury patients much better than pure kinematic models. Altogether, this panel reframed proprioception in a mechanistic context, with the hope that a better understanding of proprioception could provide a better understanding of coordinated motor control.

Despite the clear importance of somatosensation in coordinated motor control, the mechanisms of this strong influence are not well understood. However, in his early career award keynote, Andrew Pruszynski presented two remarkable experiments showcasing the surprising sophistication of the link between the somatosensory system and the control of movement. The first experiment was motivated by the simple fact that classic studies of tactile perception show large errors in describing the orientation of an edge under our fingertips. Despite this lack of perceptual acuity, humans can still perform complex, tactile-guided tasks, like clasping a necklace. Pruszynski went on to show this disparity comes from a difference in how information is used to perform the movements. In an experiment where subjects were asked to feel an edge and then orient the edge to be vertical, errors in orientation were an order of magnitude lower than found during a perceptual reporting task (Pruszynski et al. 2018). This result reflects a stark difference in somatosensory information used for motor control (which seems to be fast and precise) and the information used for perception (which is relatively slow and imprecise). In the second experiment, Pruszynski went on to present new data showing spinal reflexes based on proprioception, despite their low-level nature, can actually be quite sophisticated. Pruszynski showed in a simple arm posture stabilization task, the short latency, spinal reflex depended on whether the subject was told to resist a perturbation. Surprisingly, this reflex appeared to integrate information over multiple joints to achieve this high-level postural goal, even when the wrist was flipped around.

As one might expect, the interplay between somatosensation and motor control is not a unidirectional relationship. As discussed during many presentations, voluntary movements can also affect the transmission of somatosensory information, often in the form of suppressing sensory neural activity. This “sensory gating” is thought to aid in distinguishing between self-generated and externally imposed movements. Shubo Chakrabarti’s panel explored this idea from many different perspectives, from rodent models to humans.

Chakrabarti presented work identifying the neural pathways underlying tactile sensing in the rodent whisker system. After identifying the anatomical connections underlying the whisker system (Smith et al. 2015), Chakrabarti set up a behavioral paradigm that differentiates between passive whisker deflection (an actuated rod moves the whisker) and active whisker movement (the rat whisks against a static rod). They find that lesioning corticofugal input to the brain stem nuclei abolishes sensory gating, suggesting that the control of sensory gating is top-down.

Not all types of sensation are treated equally by gating, however. Kazuhiko Seki showed that the two main modalities underlying somatosensation, proprioception and cutaneous sensation, are gated differently by voluntary movement. While gating of proprioceptive input appeared to be large and fast, gating of cutaneous input appeared to be relatively small and slow. Additionally, Seki showed evidence that both motor and sensory cortices strongly project to the cuneate nucleus, the main brain stem nucleus for relating arm somatosensation to the cortex. Like Chakrabarti’s work, this suggests a strong influence of cortex on sensory gating.

Like the arm somatosensory system, the vestibular system is multimodal, integrating sensory information from the inner ears and proprioceptive information from the neck to maintain head and body posture. Kathleen Cullen specifically explored the idea of sensory gating in this system and how the cerebellum plays into controlling head movements. Cullen showed that vestibular signals were highly attenuated during active head movement, a classic case of sensory gating. However, this attenuation appears to occur only when sensory information matched the intended head movement: when Cullen imposed a perturbation to cause an error in head movement, this sensory gating was diminished. Cullen went on to show that neurons in the primate rostral fastigial nucleus (a deep cerebellar nucleus) explicitly encode this unexpected sensory stimulation in a way that looks like a sensory error signal (Brooks and Cullen 2013). As one might expect, this apparent error signal declines as the animal learns to deal with the imposed perturbation, possibly through updating of an internal model (Cullen and Brooks 2015).

While studying sensory gating in animal models allows for detailed recordings of fine-scaled neurophysiological signals, one consistent disadvantage is the difficulty of assessing how perception is changed during voluntary movement. James Kilner presented work focused on the difference between these two types of gating: physiological sensory gating (an actual reduction in SEP amplitude) and perceptual sensory gating (a reduction in perception of a movement). Consistent with previous studies, Kilner found significant physiological gating in response to median nerve stimulation and significant perceptual gating during a force-matching task. While late time course SEP attenuation was correlated with perceptual gating, interestingly the early parts of this attenuation did not depend on whether perception was gated during the task, suggesting a role for physiological gating other than perception. Kilner went on to suggest the early physiological gating is used for movement initiation. The first piece of evidence comes from the observation that patients with Parkinson's disease show a dramatic reduction in SEP attenuation when taken off medication (Macerollo et al. 2016). Second, Kilner found that providing sensory stimuli (via a vibrotactile motor) to medicated Parkinson’s patients actually improves performance in simple tasks. This shows that the akinetic symptoms of Parkinson’s disease can be alleviated by introducing sensory stimulation at the periphery, suggesting that at least some of the observed symptoms in the Parkinson’s could be due a failure of sensory attenuation.

CONCLUSION

Overall, advances in the field of motor control spanned several categories; however, these categories are interrelated. Developing accurate models of motor control requires a sound understanding of the neural underpinnings controlling movements as well as other neural circuits such as somatosensory feedback. Detailing the changes in neural responses from the network to the cellular levels is vital toward developing such models. Equally important is to develop models and theories based on behavioral experiments that most resemble real-world behavior rather than overanalyzing simplified laboratory tasks. As evidenced by the research highlights above, each of these topics was significantly advanced at this year’s Neural Control of Movement meeting.

GRANTS

The authors thank the Society for the Neural Control of Movement for funding to support attendance to the meeting. K. A. Mazurek was supported by grant F32NS093709 from the National Institute of Neurological Disorders and Stroke. M. Berger was supported by grant DFG RU-1847 from the German Research Foundation. T. Bollu was supported by the Cornell Mong Neurotech Fellowship. R. H. Chowdhury was supported by grant DGE-1324585 from the National Science Foundation and grants NS048845 and NS095251 from the National Institute of Neurological Disorders and Stroke. M. H. Sohn was supported by grant F32HD094552 from the National Institute of Child Health and Human Development.

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the authors.

AUTHOR CONTRIBUTIONS

K.A.M. prepared figures; K.A.M., M.B., T.B., R.H.C., N.E., I.A.K., and M.H.S. drafted manuscript; K.A.M. and M.B. edited and revised manuscript; K.A.M., M.B., T.B., R.H.C., N.E., I.A.K., and M.H.S. approved final version of manuscript.